feat(workers): free Ollama VRAM before loading Whisper on the GPU

Whisper (CT 311) and Ollama (CT 102) share one A2000. Before loading
Whisper on CUDA, ask Ollama to unload its models (GET /api/ps then POST
/api/generate keep_alive:0) and wait for the card to clear, so the GPU
load has headroom. Best-effort and stdlib-only; Ollama reloads
cooperatively, and the existing CUDA->CPU fallback covers any failure.
Toggle via OLLAMA_FREE_BEFORE_STT; endpoint via OLLAMA_URL.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

workers/void_workers/config.py

@@ -15,6 +15,12 @@ WHISPER_MODEL = env("WHISPER_MODEL", "small.en")
 WHISPER_CACHE = env("WHISPER_CACHE", "/var/lib/void/whisper-models")
 ALLOW_PRIVATE = env("VOID_INGEST_ALLOW_PRIVATE", "false") == "true"
 
+# GPU sharing: Whisper and Ollama (CT 102) share one A2000. Before loading
+# Whisper on the GPU, ask Ollama to unload its models to make room (it reloads
+# cooperatively on its next request). Best-effort; CPU fallback covers failure.
+OLLAMA_URL = env("OLLAMA_URL", "http://192.168.1.185:11434")
+OLLAMA_FREE_BEFORE_STT = env("OLLAMA_FREE_BEFORE_STT", "true") == "true"
+
 CONCURRENCY = {
     "extract.pdf":      env_int("VOID_CONCURRENCY_EXTRACT_PDF", 2),
     "extract.image":    env_int("VOID_CONCURRENCY_EXTRACT_IMAGE", 2),

workers/void_workers/gpu.py

@@ -0,0 +1,65 @@
+"""Cooperative GPU sharing with Ollama.
+
+Whisper (this worker, CT 311) and Ollama (CT 102) both pass through Z's single
+RTX A2000. Before Whisper loads on the GPU we ask Ollama to unload its models so
+there's room; Ollama transparently reloads on its next request. Everything here
+is best-effort and never raises — if Ollama is unreachable or slow, Whisper
+still tries the GPU and falls back to CPU (see model.py).
+
+Stdlib urllib only (the workers carry no `requests`/`httpx` dependency).
+"""
+import json
+import time
+import urllib.request
+
+from .log import log
+from . import config
+
+
+def _http(method, url, body=None, timeout=5):
+    data = json.dumps(body).encode() if body is not None else None
+    req = urllib.request.Request(
+        url, data=data, method=method,
+        headers={"Content-Type": "application/json"},
+    )
+    with urllib.request.urlopen(req, timeout=timeout) as r:
+        raw = r.read().decode()
+    return json.loads(raw) if raw else {}
+
+
+def loaded_ollama_models(base=None, timeout=3):
+    """Names of models Ollama currently holds in memory (GET /api/ps)."""
+    base = base or config.OLLAMA_URL
+    try:
+        data = _http("GET", f"{base}/api/ps", timeout=timeout)
+        return [m["name"] for m in data.get("models", []) if m.get("name")]
+    except Exception as e:
+        log.info("ollama_ps_failed", err=str(e))
+        return []
+
+
+def free_ollama_vram(base=None, wait_s=6.0):
+    """Ask Ollama to unload its loaded models, then wait (briefly) for the VRAM
+    to actually free. Returns the list of models it tried to unload. No-op when
+    OLLAMA_FREE_BEFORE_STT is disabled or nothing is loaded. Never raises."""
+    if not config.OLLAMA_FREE_BEFORE_STT:
+        return []
+    base = base or config.OLLAMA_URL
+    models = loaded_ollama_models(base)
+    if not models:
+        return []
+    # keep_alive:0 tells Ollama to drop the model from memory immediately.
+    for name in models:
+        try:
+            _http("POST", f"{base}/api/generate",
+                  {"model": name, "keep_alive": 0}, timeout=8)
+        except Exception as e:
+            log.info("ollama_unload_failed", model=name, err=str(e))
+    # Confirm the card is actually clear before we hand it to Whisper.
+    deadline = time.monotonic() + wait_s
+    while time.monotonic() < deadline:
+        if not loaded_ollama_models(base):
+            break
+        time.sleep(0.3)
+    log.info("ollama_vram_freed", models=models)
+    return models

workers/void_workers/model.py

@@ -32,6 +32,12 @@ def whisper_model():
         # another process sharing the card). HA portability + a shared GPU
         # mean this must degrade gracefully, never hard-fail a transcription.
         if cuda_available():
+            # Make room on the shared GPU first (best-effort; never raises).
+            try:
+                from . import gpu
+                gpu.free_ollama_vram()
+            except Exception as e:
+                log.info("ollama_free_skipped", err=str(e))
             try:
                 _whisper_model = _load_whisper("cuda", "float16")
             except Exception as e: